SCRIBE: Diagnostic Evaluation and Rich Transcription Models for Indic ASR
Researchers have introduced SCRIBE, a new diagnostic framework designed to improve automatic speech recognition (ASR) for Indic languages. Unlike traditional metrics like Word Error Rate (WER), SCRIBE categorizes errors into lexical, punctuation, numeral, and domain-entity types, offering a more nuanced evaluation. The framework also incorporates sandhi-tolerant alignment and domain vocabulary injection to better handle agglutinative languages. Alongside SCRIBE, the team has released LLM curation pipelines, benchmarks, and open-weight rich transcription models for Hindi, Malayalam, and Kannada. AI
IMPACT Enhances ASR accuracy for under-resourced Indic languages, potentially improving accessibility and usability.